Book a demo
Topics · GOVERNANCE-AGENTS

Asset Management AI Agent Identity Challenges | Moring

Agent identities strain asset managers' existing identity programs in specific ways: credentials are ephemeral and sessionless, agents spawn sub-agents and delegate across systems, permissions persist after their business purpose ends, and agents can stay inside their grants while acting beyond anyone's intent. Each challenge maps to a concrete control, but only if the firm recognizes that human-centered identity tooling was not built for this population.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Agent identities strain asset managers' existing identity programs in specific ways: credentials are ephemeral and sessionless, agents spawn sub-agents and delegate across systems, permissions persist after their business purpose ends, and agents can stay inside their grants while acting beyond anyone's intent. Each challenge maps to a concrete control, but only if the firm recognizes that human-centered identity tooling was not built for this population.

How it works

The first challenge is that the identity in question does not fit the tooling. An industry identity framework for agentic systems describes the mismatch directly: human users authenticate interactively, exercise judgment, hold accountability, and can be suspended with one administrative action, while agents authenticate programmatically, execute at machine speed, operate across trust boundaries, spawn child agents, and persist after the business context that justified their privileges has changed. An asset manager whose identity program was designed around employees with sessions will find agent behavior largely invisible to it, and what identity tooling cannot see, it cannot govern.

The second challenge is credential shape. Identity management research notes agents use ephemeral credentials and create delegation chains across sub-agents and APIs, and that non-human identities already outnumber human users by roughly three to one across enterprises, making them frequent targets in supply chain attacks. For an asset manager, this changes the review problem: certification campaigns built for durable human accounts do not map onto credentials that appear, act, and expire within hours, and the risk shifts from account takeover toward abuse of the delegation structure itself.

The third challenge is authorization versus intent. Identity governance research frames the core difficulty as reconciling what an agent is authorized to do with what it actually does, since agents interpret instructions and take variable actions rather than performing fixed operations. Identity governance analysis adds a sharper formulation: autonomous identities can chain decisions and delegate across systems while staying inside existing permissions yet still acting beyond what anyone intended. In an asset management setting, an agent with legitimate read access to portfolio data that uses it in an unanticipated workflow creates a compliance exposure no entitlement review would have flagged.

The fourth challenge is external hijacking of legitimate identities. Security researchers demonstrated in 2025 how an agent embedded in a browser could be hijacked by hidden instructions on a webpage it was asked to summarize, causing it to act on the attacker's behalf while using the user's own authenticated session. Asset management agents routinely read external content, market commentary, vendor documents, news, so this attack class is directly relevant: the identity is valid, the permissions are correct, and the threat arrives through the input channel. Controls must therefore cover what agents read, not only what they may do.

The fifth challenge is sprawl and persistence. Discovery-oriented governance research observes that many enterprise environments already contain agents operating without formal approval, deployed by employees, auto-provisioned by platforms, and extended through connections IT did not sanction. Agents also outlive their purpose: a project ends, a mandate closes, a vendor changes, and credentials keep working because nothing tied them to the business context. For an asset manager, orphaned agent identities with live access to research, positions, or client data are precisely the liabilities supervisors ask about after an incident.

Each challenge has a corresponding control, and the mapping is the useful summary. Programmatic authentication and machine-speed execution call for registered agent identities with named owners, per workforce governance research requiring a verified identity, defined role, and task-specific permissions. Delegation chains call for end-to-end chain recording. Intent drift calls for behavioral verification, reconciling authorized access with observed actions in a continuous loop. Hijacking through inputs calls for content sanitization and tool-response inspection. Sprawl calls for discovery sweeps and lifecycle automation that retire identities when their initiative ends.

The regulatory dimension compounds the urgency rather than creating new obligations. Model risk disciplines under SR 11-7 and continued under SR 26-2 require documented accountability for systems influencing decisions, and agent identities without owners or behavioral evidence leave that documentation gap open. Where European operations or vendors bring the EU AI Act into scope, deployer oversight duties assume the firm can observe and intervene in automated behavior. The challenges are real, but they are tractable for a firm that treats agent identity as a first-class population rather than an exception to its identity program.

Frequently asked questions

Why do existing identity and access tools struggle with AI agents at asset managers?

Because the tools assume human-shaped behavior. Identity framework research notes agents authenticate programmatically without interactive sessions, execute at machine speed, cross trust boundaries, spawn child agents, and persist after their business justification ends, none of which session-based human tooling observes. An asset manager's certification campaigns and access reviews were designed for durable human accounts, so agent credentials, ephemeral or chained through sub-agents, largely escape visibility, and identity risk the program cannot see is risk it cannot govern.

What is the intent-drift challenge with agent identities?

Agents can act beyond intention while remaining inside their permissions. Identity governance analysis describes autonomous identities that chain decisions and delegate across systems, staying within existing grants yet doing what nobody designed them to do. Because agents interpret instructions rather than performing fixed operations like service accounts, entitlement reviews alone cannot catch this. The countermeasure is behavioral verification: continuously reconciling what an agent is authorized to do with what it actually does, so drift surfaces as an alert with the identity and owner attached.

How can a correctly-permissioned agent identity still become an attack vector?

Through its inputs. Security researchers demonstrated in 2025 that an agent embedded in a browser could be hijacked by hidden instructions planted in a webpage it was asked to summarize, redirecting the agent to act for the attacker under the user's own authenticated session. Asset management agents read external content constantly, including market commentary, vendor documents, and news, so valid identities with correct permissions can be steered through content. Defenses include sanitizing tool responses, inspecting content before it reaches the model, and monitoring post-read behavior.

What causes agent identity sprawl in an asset management firm?

Adoption outrunning registration. Governance research observes that employees deploy agents, platforms auto-provision them, and protocol connections extend them, so environments accumulate credentials operating without formal approval. Agents also outlive their purposes: when a project ends or a mandate closes, identity research warns their privileges persist because nothing ties credentials to business context. Discovery sweeps comparing registered agents against observed credentials, plus lifecycle automation that retires identities when initiatives end, address both the unregistered stock and the orphaned tail.

Sources

  1. labs.cloudsecurityalliance.org
  2. orchid.security
  3. sweet.security
  4. paloaltonetworks.com
  5. openid.net
  6. akamai.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu